中山大学学报自然科学版 ›› 2011, Vol. 50 ›› Issue (2): 6-10.

• 研究论文 • 上一篇    下一篇

基于概率图模型的图像纹理模型

杨 关,冯国灿,陈伟福,罗志宏   

  1. (中山大学数学与计算科学学院,广东省计算科学重点实验室,广东 广州 510275)
  • 收稿日期:2010-03-15 修回日期:1900-01-01 出版日期:2011-03-25 发布日期:2011-03-25
  • 通讯作者: 冯国灿

Texture Models Based on Probabilistic Graphical Models

YANG Guan, FENG Guocan, CHEN Weifu, LUO Zhihong   

  1. (School of Mathematics and Computational Sciences, Sun Yatsen University,Guangdong Province Key Laboratory of Computational Science,Guangzhou 510275, China)
  • Received:2010-03-15 Revised:1900-01-01 Online:2011-03-25 Published:2011-03-25

摘要: 纹理作为一种视觉特征,它广泛应用于图像分析。概率图模型由于其自身特点可以很好地描述纹理。高斯图模型结构可根据局部马尔科夫性和高斯变量的条件回归之间的关系来学习。高斯图模型可用一个邻域系统、一个参数集和一个噪声序列表示。利用惩罚正则化方法,可以选择高斯图模型的邻域并估计参数,然后提取纹理特征进行纹理合成和分类。实验结果显示基于高斯图模型的纹理特征更加有效。

关键词: 高斯图模型, 模型选择, 惩罚正则化, 纹理合成, 纹理分类

Abstract: Texture is one of the visual features playing an important role in image analysis. Many applications have been discovered using texture models.Probabilistic graphical models Science,are promising tools for constructing texture models.The problem of learning the structure of GGM for texture classification is addressed. GGM are characterized by a neighborhood, a set of parameters, and a noise sequence due to the connection between the local Markov property and conditional regression of a Gaussian random variable. By use of the methods of model selection to choose an appropriate neighborhood and estimate the unknown parameters for modeling GGM, neighborhood selection and parameter estimation are conducted simultaneously. And then new texture features based on GGM for texture synthesis and texture classification are extracted.Experimental results show that adaptive Lasso estimators are more effective.

Key words: Gaussian graphical models, model selection, penalty regularization, texture synthesis, texture classification

中图分类号: